improv.sh

Most AI coding agents rediscover the task on every run, interviewing you across multiple turns before producing anything useful. improv addresses this with a task harness: a single @im or /im call compiles a vague request into a structured spec containing objective, context, constraints, deliverables, acceptance criteria, and shell validation commands — then passes that spec to the agent so implementation starts on turn one. The loop closes not with LLM self-report but with real exit-code judges; the agent implements, the judges run, failures feed back in, and the cycle repeats until exit code 0.

The differentiating feature is the loop infrastructure. The vendor describes slice-by-slice execution checklists — inspect, implement, verify, report — combined with task memory persisted under .improv/tasks/. Agents forget context between sessions; the task lifecycle ID and status tracking mean the harness does not. Repo-aware context detection pulls package scripts, test commands, and agent instruction files from the workspace automatically, so the spec reflects your actual stack rather than a generic template.

improv installs locally via a curl script with no API key required, and the Chrome extension sideloads the same engine into browser-based chat interfaces including Claude Code, Codex, and Cowork. VS Code and Cursor users get a Marketplace extension that lets them diff, apply, and run judges without leaving the editor. The 920-skill library grows via a daily self-improving search-score-adopt pipeline, which the vendor states runs automatically. There is no API surface documented on the page, so integrating improv into a custom CI pipeline or programmatic workflow is not supported by the current architecture.